The framing here is tidy: a video is a world plus an event stream. The world is the slow stuff, the room, the people, the ambient sound, a voice's timbre. The event stream is everything moving through it, speech, actions, scene changes. Split it that way and you get a clean pretraining goal: given the world and whatever just arrived, predict how it responds next, in real time.

Wan-Streamer v0.3 puts this to work on full-duplex audio-visual interaction, where the agent's own speech and behavior are the event stream, so the model maps live input to spoken and physical actions, much like a vision-language-action system. It holds the previous version's operating point: 640 by 368 video at 25 fps, roughly 200 ms of model-side latency and about 550 ms end to end under a modest network budget.

This is a system-and-framing report, so the paper is where the training details and honest limits should live.